Does this model predict absolute depth?
I am interested in a model able to predict the absolute depth in meters. So far I am not able to figure out how to obtain this nor am I sure that this is possible at all. Can someone shed light on this issue.
When I use the code, the predicted pixel values are high for close objects and small for far away objects. I found discussions that indicate that the model produces inverse depths and that scaling and shift factors are required but I can't figure them out.
Answering myself: This model is trained on an immense amount of training data. This is made possible by only leveraging the relative depths in the data sources instead of absolute metric depths values. Many of the data sources do not provide absolute depth to begin with (3D movies). Thus every prediction is relative [inverse] depth scaled to the depth range of each image (d_min -> 0, d_max -> 255). For the comparison to state of the art methods, this depth range is taken from the ground truth images.
Indeed, if you're interested in an absolute depth estimation model check out ZoeDepth: https://huggingface.co/spaces/shariqfarooq/ZoeDepth